Maria Sindeeva

RNAs are biological polymers that, like proteins, can fold into structures and perform many important enzymatic or structural functions in cells. While Alphafold and related methods can predict the structures of folded proteins, no existing tool can as yet reliably predict the structures of folded RNAs. A critical bottleneck is the difficulty to find similar (homologous) RNA sequences in the databases. In this project, we will develop new ways to represent RNA structures by generative training of transformer networks similar to LLMs (e.g. MSA pairformers). RNA sequence similarity searches will then utilize these representations instead of the actual nucleotide sequences to greatly boost the sensitivity of structure searches. (See our recent tool Foldseek, van Kempen et al, Nature Biotechnol. 2024). The better searches should allow us to build more informative multiple sequence alignments (MSAs), with which the MSA Pairformer should produce more informative embeddings, with which RNA structure modules should have a better chance to predict correct RNA structures de novo.